Odyssey whdload (demo) freeze under winuae

[vagrant]

So early in the demo during the part where a ship flies across a planets surface, the ship freezes and landscape continues forever..
I've tried (seemingly) everything but not matter what the demo always bugs at this point.

I know there's a few other threads addressing the issue but the OP's have just given up and gone onto floppy versions. Apparently works ok on real amiga but not winuae.

[ascp]

File a bugreport at http://www.whdload.de/demos/Alcatraz_Odyssey.html so that the authors may notice it.

[Toni Wilen]

With and/or without JIT? (2.6 betas have some JIT specific exception tweaks and I think this demo used division by zero exceptions during that part)

[vagrant]

Yep it happens with JIT enabled or disabled, cycle exact enable/disable also tried blitter options, 'more compatible' etc. (I even tried all 3 versions of the demo)

[Toni Wilen]

I checked it, I remembered correctly, it assumes undocumented DIVS behavior when dividing by zero (V flag getting cleared).

68020 sets V and UAE uses 68020 behavior if CPU is 68020+. 68060 (perhaps 68040 too) seems to again clear V.

So this demo won't work on 68020 (real or emulated) but it may work on real 68030+ (or maybe only 68040+).

[Wepl]

This could probably fixed in the whdload slave.
Can you show me a code snippet and address where the bad code is located?

[Toni Wilen]

Originally Posted by Wepl

This could probably fixed in the whdload slave.
Can you show me a code snippet and address where the bad code is located?

00082F9C 0c86 0000 7fff           CMP.L #$00007fff,D6
00082FA2 6f06                     BLE.B #$00000006 == $00082faa
00082FA4 5245                     ADD.W #$00000001,D5
00082FA6 e286                     ASR.L #$00000001,D6
00082FA8 60f2                     BT .B #$fffffff2 == $00082f9c
00082FAA 8fc6                     DIVS.W D6,D7
00082FAC 6806                     BVC.B #$00000006 == $00082fb4
00082FAE e287                     ASR.L #$00000001,D7
00082FB0 5345                     SUB.W #$00000001,D5
00082FB2 60f6                     BT .B #$fffffff6 == $00082faa

BVC.B at 82FAC should jump (V=0) if DIVS tried to divide by zero. (Divide by zero exception does nothing, it points directly to RTE)

(and if someone is bored enough, 68030 and 68040 needs divs/divu divide by zero flag test)

[Wepl]

so clearing the V flag in the exception handler should sufficient I think
I will check behavior on 30/40...

[Toni Wilen]

I just noticed 68020 DIVS and DIVU has very different flag bebavior when dividing by zero, DIVU is correct but DIVS isn't correct in UAE.

I think it is better to not "fix" this until all CPU models are tested. There may not be anything to fix.. (I don't want fixes for bad emulation, they only hide possible future problems)

68020 and 68030: Signed divide: Z=1, NVC=0. Unsigned: V=1, N=dst<0, Z=!N. (Weird, isn't it?)

68060: Clear C. All other flags are preserved.

EDIT: I was wrong, demo works if CPU clears V or does not modify it. It only gets stuck if V is set by the CPU.

EDIT2: A1200 + 68020 and 68060 tested and space ship part confirmed working.

EDIT3: 68030 tested. Exact same as 68020 (was not surprising). I'd assume 68040 does same as 68060.

-> No need for slave fixes.

[Wepl]

wrote a test slave:

MC68060 UAE=N
divu X=X N=X Z=X V=X C=0
divs X=X N=X Z=X V=X C=0

MC68040 UAE=N
divu X=X N=X Z=X V=X C=0
divs X=X N=X Z=X V=X C=0

MC68030 UAE=N
divu X=X N=1 Z=0 V=1 C=0
divs X=X N=0 Z=1 V=0 C=0

MC68020 UAE=N
divu X=X N=1 Z=0 V=1 C=0
divs X=X N=0 Z=1 V=0 C=0

question remains if results on 20/30 depends on the value of the dividend

[Toni Wilen]

Thanks. 68040 = 68060, everything is fine

Originally Posted by Wepl

question remains if results on 20/30 depends on the value of the dividend

It does. "dst" in my previous test was dividend. (dividend < 0: N=1 Z=0, dividend >= 0: N=0 Z=1)

Documentation also says overflow makes N and Z undefined. More testing to do, later..